In late September, Shopify extended its WebMCP support to checkout, which means browser-based AI agents can now read a cart, edit order details and submit a purchase with the buyer's authorization, Shop Pay included, without screen-scraping. It is rolling out to all eligible merchants, it is on by default, and merchants can opt out. Amazon and Adidas have reportedly gone the other way and block agents from buying on a user's behalf. If you sell online, you now have a decision you probably have not made.
What agentic checkout actually is
Until now, an AI agent trying to buy something on your site had to behave like a clumsy human: read the rendered page, guess which button to click, fill fields it inferred. It broke constantly and looked like bot traffic because it effectively was.
WebMCP changes the mechanism. Your store exposes structured tools an agent can call directly, so the agent is not interpreting a page, it is using a defined interface. Shopify's storefront and checkout tooling sit on its Universal Commerce Protocol. The practical upshot is that agent purchases become reliable rather than flaky, which is exactly why the volume question stops being hypothetical.
The buyer still has to approve before an order is submitted. That is the guardrail, and it matters for the liability discussion below.
The case for leaving it on
If customers start shopping through assistants and your store is one of the ones that works, you are in the consideration set and your competitors who opted out are not. Meta's Muse and Instinct already have direct Shopify partnerships. This is the same shape as the search transition: being machine-readable early was cheap, and being invisible later was expensive. That is the argument we made about getting cited by AI assistants, and checkout is the same logic applied to the transaction instead of the answer.
The case for turning it off, honestly
Three real reasons, none of them hysterical.
Your fraud tooling was trained on humans. Risk scoring leans on signals like typing cadence, mouse movement, session length and device fingerprint. An agent produces none of the human ones. Expect more false positives, more manual review, or rules that quietly pass things they should not.
The dispute framework is unsettled. The trade press is blunt that the industry has not agreed standards for authorization, liability and dispute handling in agent-completed transactions. If a buyer says their agent bought the wrong thing, or bought without real authorization, the chargeback lands on you while the rules are still being written.
You may not be ready to serve the order. An agent that buys in thirty seconds does not read your shipping notes, your made-to-order lead time or your "contact us for bulk pricing" line. For some businesses that is a support problem, not a sale.
What quietly breaks
- Abandoned-cart email. An agent that browses and leaves is not a shopper you should chase, and an agent-submitted email address may not be a person who wants your flow.
- Analytics and attribution. Agent sessions distort conversion rate, time on page and channel attribution. If you make spend decisions on those numbers, you want to be able to segment agent traffic out.
- CAPTCHA and bot rules. Any defence that assumes automation equals abuse will now block legitimate, authorized purchases. Those rules need revisiting rather than removing.
- Loyalty and promo logic. One-per-customer offers, referral credit and points assume an identifiable human. Decide how they behave when an agent transacts.
- Address and tax edge cases. Agents fill fields confidently. Validate server-side rather than trusting the input.
What to actually do
- Find the setting and make a decision. Even leaving it on is fine. Inheriting a default without knowing is not.
- Segment agent traffic in analytics before volume arrives, so your baseline is not polluted retroactively.
- Re-read your bot rules and separate "automation" from "abuse". They are no longer the same thing.
- Make your product data machine-readable. Clear titles, real specifications, accurate stock and unambiguous pricing. An agent comparing three stores picks the one whose data it can actually parse, which makes this a merchandising task as much as a technical one.
- Watch your first agent orders individually. Ten manual reviews will tell you more than any policy you write in advance.
- Check your payment terms. Ask your processor, in writing, how they treat agent-completed transactions for chargeback purposes. Most will not have a clean answer yet, and that answer is itself information.
- Keep the duller things working. Payment-page hygiene and patching matter more than any of this, because a skimmed checkout is a worse outcome than a missed agent sale.
What this means for your business
Agentic commerce is at the stage where the cost of engaging is low and the cost of ignoring is unknown. That argues for a deliberate, small position: leave the door open, instrument it so you can see what comes through, and fix the fraud and analytics assumptions that quietly assume a human. The businesses that will struggle are not the ones that chose wrong, they are the ones that never knew the setting existed and spend next year wondering why their conversion data stopped making sense.